Nitin Khanna

I was about to write something similar. I have spent 2 months building a tool that now has over 400K lines of code. Every time it finishes and tests run, I ask it are you lying, its not working, and it adds more code. This is CoT applied at it best. When I asked to delete unused code it deleted close to 100K :D

GPT 5.6 Sol Ultra worked in /goal mode for two days and wrote 200k lines of code, when gently questioned it said 80% was a waste of time and reverted it all. There's something deeply fundamental that's still missing. They understand the how but never the why, no telos.

@jonkomet Interesting you say that, because I started hitting these a lot more now

ok now i can say, that GPT 5.6 has resolved this token problem with efficiency and i am no longer experiencing bad usage limit issues🔥

@sama Yeah it has been a crazy ride with fable. Sometimes, it hard to follow along.

30% of the cost was on fable at these levels of usage?

@codetaur Wait to see if we need to review your prompt. I got it 4 times already

sama I never wanna see this shit again

@petergyang @rezoundous For me high was able to close goals which were running for almost a week with no end in sight

@rezoundous Medium for me high shreds my tokens

After a day of trying GPT-5.6, I've found the most balanced for me to be GPT-5.6 Sol High. What about you guys?

@AIandDesign ✅☑️👍

Unpopular opinion: The better the engineer, the less it matters what LLM they’re using to code with.

I have been testing Gemma 4 for summarization on my local MacBook over the last few days. I am impressed by its capabilities. One tool I built is to summarize YouTube channels that I would like to go through in a focused manner. While the summarization tasks are still going on, I thought it might be at a good enough state to share the ai-dot-engineer conference summaries with you: tldr.productexperts.cc/?channel=aidot…

@alightinastorm Welcome to software engineering, my friend. The essence is the same; the tools have changed.

my preferred way of vibe coding is to first do very broad horizontal coding (get the most minimal version of any feature done) to properly understand the domain and feasibility of the project once that is done, i control the focus on the horizontals and AI turns into a deep vertical coding specialist attention is finite. for the ai and for you. if you keep pumping broad instructions across the whole app it'll never produce anything meaningful

@rezoundous I think, @SpaceXAI may come out as a final winner here, just because they have compute and can control the levers much better.

With every model upgrade, the usage limits downgrades.

@nateparrott It feels like start of an affair, but then 2 hours in you are like can I quit already :D

folks who’ve tried Claude Design — how was the experience? what works well and what could be better?

@DavidOndrej1 For me, it is fixed. It was able to complete several loops that had been stuck for days for me. And I was not seeing meaningful progress. But for everything of value, it seems like they are reviewing the prompts. I am not happy with it. Planning to switch.

does OpenAI have half of the AI influencers on payroll? because I'm not impressed with GPT 5.6 at all...

@kcosr I have had a similar experience with a couple of these models. I am seeing a growing tendency to shove functionality despite guidance not to do so. I think this leads to over-engineered code.

Just deleted 4K LOC of completely unnecessary, over-engineered (and less performant) C++ coded added by GPT-5.6 Sol xhigh. Not a good start. I did have a good experience with Terra earlier today.

I think, in a lot of ways, I have mixed feelings too; it tends to jam a lot of stuff into a single experience despite being told not to. Also, they are triggered this weird wait till we review flow which makes me feel uncomfortable, they are publically saying that we are looking at what you are working on.

@DavidOndrej1 Why not?

first impressions of GPT 5.6 are not good...

@jack Now that we have outsourced intelligence, how long does it take for us to understand the dynamics? The deterioration of the human brain is on an accelerated curve.

i’ve shifted from telling agents what to do, to asking them what to do, and pulling the best thread.

@dayonefoundry A serious product takes time. I think what you are referring to is mvp for mvp

Me: How long will this project take. Claude: About 12-18 months. Me: But you're an AI Claude: Ah yes if you leverage AI Coding then the estimate would be 6-9 months. Me: But you're an AI Claude: Even with AI, there are many aspect of coding that take time. AI can only do much to accelerate coding but it is not a magic tool. Now that we have a realistic estimate, would you like to get started? Me: Yes *3 hours later* Claude: Ok I'm done.

@amritwt Not there yet but you are directionally correct

Never thought I’d say this but Grok 4.5 is much better than whatever Anthropic offers right now

@anupamrjp 100%, but it's expensive as hell for anything meaningful, is my first response. If there is a law of conservation for applied intelligence patterns, I think Sol and Fable are the same. One costs more; the other takes more time. Currently, I prefer time over cost.

Be honest: is Fable actually better than Codex, or is the hype just louder?

@iuditg The UX is great.

How is Grok 4.5 and how are the limits? Looking for some honest feedback.

@majidmanzarpour They are one product, I think. It's the domain spread that is interesting for these companies right now, more than user satisfaction to justify valuation.

Instead of separate ChatGPT Chat/Work/Codex, and Claude Chat/Cowork/Code, why not just make them 1 single interface? It's confusing UX for average users. The Codex & Claude Code harnesses can do everything the Chat/Work harnesses do and are very reliable. Figure out what to do based on the user intent, if the user asks for a coding tasks, do code. If the user asks for a work task, do work. If the user wants to chat, do chat.